영역별 대역간 양방향 예측과 확장된 SPIHT를 이용한 다분광 화상데이터의 압축

김승진,반성원,김병주,박경남,김영춘,이건일,Kim, Seung-Jin,Ban, Seong-Won,Kim, Byung-Ju,Park, Kyung-Nam,Kim, Young-Choon,Lee, Kuhn-Il 대한전자공학회 2002 電子工學會論文誌-SP (Signal processing) Vol.39 No.5

본 논문에서는 웨이블릿 영역에서 각 부밴드에 대한 영역별 대역간 양방향 예측과 확장된 SPIHT (set partition in hierarchical trees)를 이용한 효율적인 인공위성 다분광 화상데이터의 압축 방법을 제안하였다. 이 방법에서는 가시광선 영역과 적외선 영역에서 다른 대역과 분광적 상관성이 큰 대역을 기준대역 (feature band)으로 각각 결정하고, 이 대역들에 대해 웨이블릿 변환 (wavelet transform, WT)을 행한 후 SPIHT를 행하여 부호화함으로써 대역내 (intraband) 중복성을 제거한다. 기준대역과 대역간 상관성이 큰 예측대역 (prediction band)들에 대해서는 웨이블릿 변환을 행한 후, 각 대역의 기저밴드의 대역별 특성을 이용하여 영역분류를 하고, 각 부밴드에 대한 영역별 대역간 양방향 예측 (classified interband bidirec- tional prediction)을 행함으로써 대역간 (interband) 중복성을 제거하여 압축 효율을 향상시킨다. 또한 확장된 SPIHT의 부호화 효율을 높이기 위해 예측오차의 최대값에 따라 재배열된 대역들에 대해 확장된 SPIHT를 행하여 예측오차를 부호화함으로써, 예측에 따른 오차를 보상하여 화질을 향상시킨다. 실제 다분광 화상데이터에 대한 모의 실험을 통하여 제안한 방법의 부호화 효율이 기존의 방법에 비하여 우수함을 확인하였다. In this paper, we proposed the effective multispectral image compression method using CIBP(classified interband bidrectional prediction) and extended SPIHT(set partition in hierarchical trees) in wavelet domain. We determine separately feature bands that have the highest correlation with other bands in the visible range and in the infrared range of wavelengths. Feature bands are coded to remove the spatial redundancy with SPIHT in the wavelet domain. Prediction bands that have high correlation with feature bands are wavelet transformed and they are classified into one of three classes considering reflection characteristics of the baseband. For Prediction bands, CIBP is performed to reduce the spectral redundancy. for the difference bands between prediction bands and the predicted bands, They are ordered to upgrade the compression efficiency of extended SPIHT with the largest error magnitude. The arranged bands are coded to compensate the prediction error with extended SPIHT. Experiments are carried out on the multispectral images. The results show that the proposed method reconstructs higher quality images than images reconstructed by the conventional methods at the same bit rate.

양방향 장단기 기억과 시계열 교차검증을 이용한 철근 가격의 장기예측에 관한 연구

이용성(Lee, Yong-Seong),김경환(Kim, Kyung-Hwan) 대한건축학회 2022 대한건축학회논문집 Vol.38 No.5

This study proposes a long-term prediction method of rebar price using deep learning techniques such as a bidirectional long and short-term memory (Bi-LSTM), a recursive method, and a time series cross-validation. Among recurrent neural network (RNN) models, Bi-LSTM provides the best prediction performance for small time series data such as monthly rebar price when applied. The recursive method uses the short-term prediction result as an input value for predicting the next time point data, which can repeatedly be used for making long-term predictions. Time-series cross-validation enables more stable prediction accuracy by enhancing learning that may be lacking in small time series data. By applying these deep learning techniques, this study predicts the monthly rebar price for up to 5 months and compares it with the previous study. As a result, it has been found that the average accuracy increases, and the deviation of the predicted values decreases.

영역별 양방향 예측과 KLT를 이용한 인공위성 화상데이터 압축

김승진,김태수,박경남,김영춘,이건일,Kim Seung-Jin,Kim Tae-Su,Park Kyung-Nam,Kim Young-Choon,Lee Kuhn-Il 대한전자공학회 2005 電子工學會論文誌-SP (Signal processing) Vol.42 No.1

웨이블릿 영역에서 영역별 양방향 예측, KLT (Karhunen-Loeve transform)/sup [13]/, 및 3차원 SPIHT (set partition in hierarchical trees)/sup [1]/를 이용한 인공위성 화상데이터의 부호화 방법을 제안하였다. 가시광선 영역과 적외선 영역에서 선택된 기준대역 (feature band)에 대하여 SPIHT를 행하여 부호화함으로써 대역내 (intraband) 중복성을 제거한다. 기준대역을 예측대역(prediction band)들에 대해서는 웨이블릿 변환 (wavelet transform)을 행한 후, 빛의 반사 및 역의 방사에 따라 대역별 특성이 다름을 이용하여 영역분류를 하고 영역별 양방향 예측 (classified bidirectional prediction)을 행함으로써 대역간 (interband) 중복성을 제거한다. 원 인공위성 화상데이터와 부호화 된 인공위성 화상데이터 사이의 오차값으로 구성된 오차대역 (residual band)들에 대하여 KLT를 행함으로써 대역간 중복성이 제거되고 계수값들은 고유치의 크기에 따라서 분광적으로 정렬됨으로써 3차원 SPIHT의 부호화 효율을 향상시킨다. 인공위성 화상데이터에 대한 모의실험을 통하여 제안한 방법의 부호화 효율이 기존의 방법에 비하여 우수함을 확인하였다. We propose an effective Landsat TM image compression method using the classified bidirectional prediction (CBP), the classified KLT and the SPIHT. The SPIHT is used to exploit the spatial redundancy of feature bands selected in the visible range and the infrared range separately. Regions of the prediction bands are classified into three classes in the wavelet domain, and then the CBP is performed to exploit the spectral redundancy. Residual bands that consist of difference values between the original band and the predicted band are decorrelated by the spectral KLT Finally, the three dimensional (3-D) SPIHT is used to encode the decorrelated coefficients. Experiment results show that the proposed method reconstructs higher quality Landsat TM image than conventional methods at the same bit rate.

Application of Informer for time-series NO2 prediction

Hye Yeon Sin(신혜연),Minchul Kang(강민철),Joonsung Kang(강준성) 한국컴퓨터정보학회 2023 韓國컴퓨터情報學會論文誌 Vol.28 No.7

본 논문에서는 딥러닝 시계열 예측 모형을 평가한다. 최근 연구에 따르면 이 모형은 ARIMA와 같은 기존 예측 모형보다 성능이 우수하다고 결론짓는다. 그 중 히든 레이어에 이전 정보를 저장하는 순환 신경망이 이를 위한 예측 모형 중 하나이다. 네트워크의 그래디언트 소실 문제를 해결하기 위해 LSTM은 데이터 흐름의 반대 방향으로 숨겨진 레이어가 추가되는 BI-LSTM과 함께 순환 신경망 내부의 작은 메모리로 사용된다. 본 논문은 서울의 2018년 1월 1일부터 2022년도 1월 1일까지의 NO2 자료에 대해 Informer의 성능을 LSTM, BI-LSTM, Transformer와 비교하였다. 이에 실제 값과 예측값 사이의 평균 제곱근 오차와 평균 절대 오차를 구하였다. 그 결과 Test 데이터 (2021.09.01.~2022.01.01.)에 대해 Informer는 다른 방법에 비해 가장 높은 예측 정확도 (가장 낮은 예측 오차: 평균 제곱근 오차: 0.0167, 평균 절대 오차: 0.0138)를 보여 타 방법에 비해 그 우수성을 입증하였다. Informer는 당초 취지와 부합되게 다른 방법들이 갖고 있는 장기 시계열 예측에 있어서의 문제점을 개선하는 결과를 나타내고 있다. In this paper, we evaluate deep learning time series forecasting models. Recent studies show that those models perform better than the traditional prediction model such as ARIMA. Among them, recurrent neural networks to store previous information in the hidden layer are one of the prediction models. In order to solve the gradient vanishing problem in the network, LSTM is used with small memory inside the recurrent neural network along with BI-LSTM in which the hidden layer is added in the reverse direction of the data flow. In this paper, we compared the performance of Informer by comparing with other models (LSTM, BI-LSTM, and Transformer) for real Nitrogen dioxide (NO2) data. In order to evaluate the accuracy of each method, mean square root error and mean absolute error between the real value and the predicted value were obtained . Consequently, Informer has improved prediction accuracy compared with other methods.

양방향 LSTM과 데이터 조합탐색 및 딥러닝 관련 기법을 활용한 철근 가격 단기예측에 관한 실험적 연구

이용성,김경환 한국건설관리학회 2020 한국건설관리학회 논문집 Vol.21 No.6

This study presents a systematic procedure for developing a short-term prediction deep learning model of rebar price using bidirectional LSTM, Random Search, data combination, Dropout. In general, users intuitively determine these values, making it time-consuming and repetitive attempts to explore results with good predictive performance, and the results found by these attempts cannot be guaranteed to be excellent. With the proposed approach presented in this study, the average accuracy of short-term price forecasts is approximately 98.32%. In addition, this approach could be used as basic data to produce good predictive results in a study that predicts prices with time series data based on statistics, including building materials other than rebars.

양방향 LSTM을 활용한 전력수요 데이터 예측 기법 연구

고상준,윤호영,신동명 한국소프트웨어감정평가학회 2018 한국소프트웨어감정평가학회 논문지 Vol.14 No.1

Power demand prediction is used to detect anomalies such as machine malfunction and power off when the measured power amount deviates from the predicted power amount by a certain range in the industrial sector. Power data is difficult to predict because it is time series data with strong seasonality and trend. In this paper, a power prediction model based on bi-directional LSTM algorithm, which is a deep learning based algorithm and the most popular algorithm in recent years, has been established along with LSTM. The model was learned, evaluated, and verified using four years of power data.

비선형 모델 예측 제어를 활용한 양방향 쿼드로터의 동일한 평탄 출력 궤적을 추종하는 비행 자세 사이의 전환

김병준(Byeongjun Kim),김현진(H.Jin Kim) 한국항공우주학회 2023 한국항공우주학회 학술발표회 논문집 Vol.2023 No.11

BRDF를 고려한 적외선 신호의 반사 성분 고속 연산기법에 관한 연구

김동건,한국일,최준혁,최순호,김태국,Kim, Dong-Geon,Han, Kuk-Il,Choi, Jun-Hyuk,Choi, Soon-Ho,Kim, Tae-Kuk 한국군사과학기술학회 2017 한국군사과학기술학회지 Vol.20 No.1

This paper is a part of developing a computer code that can be used to generate synthetic IR images by calculating the outgoing infrared signal from objects. To predict the reflected component that is a part of the outgoing IR signal, such as those components reflected from the target surface by the solar and sky irradiations, it is necessary to calculate the complicated BRDF values for considering the directional surface reflection characteristics. Since the calculation of reflectance using the BRDF requires a large amount of computation time due to the hemispherical integral term, it is frequently restricted in applying for a real-time prediction of IR signal. In this research, the simplified method for calculating IR reflected component has been proposed by replacing the integral terms into two parts, a directionally uniform component and a step function representing the specular component, to reduce computation time. The proposed method is proved to result in very fast calculation of the BRDF (up to 600 times faster calculations) for most of the surfaces with minimal loss of the accuracy.

HSKD: Hidden State Knowledge Distillation을 이용한 배터리 SoC 예측 모델 경량화

강수혁(Soohyeok Kang),박규도(Guydo Park),심동훈(Donghoon Sim),조선영(Sunyoung Cho) 한국자동차공학회 2023 한국자동차공학회 학술대회 및 전시회 Vol.2023 No.11

Knowledge Distillation (KD) is one of the representative methods for AI model compression, where a student model learns by imitating the output of a teacher model. The student model has a smaller network than the teacher model, which can reduce inference time and save memory. This method should be applied for efficient AI model inference in limited computing environments, such as the vehicle controller. In this paper, we applied the Hidden State Knowledge Distillation (HSKD) method to a Bi-LSTM (Bidirectional Long Short Term Memory) model for predicting the State of Charge (SoC) of an electric vehicle battery. This model predicts the SoC 5 minutes ahead using the SoC of the past 5 minutes. In the experiment, we selected a teacher model with a hidden size of 1,024, which showed the highest accuracy, and compared the performance of hidden state knowledge distillation and general knowledge distillation models for models with a hidden size smaller than 1,024. And, we measured the inference time of the compressed models on controllers equipped with ARM Cortex-A53. As a result, the model with a hidden size of 32 had a loss of 0.008 in terms of R2 score compared to the teacher model, but the inference time was reduced by approximately 20.1x and the file size was compressed by 750.6x from 33,028 [KB] to 44 [KB].